Speech recognition evalualtion must test where voice agents actually work, which is phone calls between two people.

Voice Arena’s new benchmark tests speech-to-text models on real calls made on everyday phones, in real rooms, with real background sound.

Two speakers make the job much harder for the model. People talk over each other, interrupt, drop in short “yeah” and “okay” replies, and the phone line compresses the audio. A support or sales agent deals with all of this every day.